iknownothow

But don't input embeddings need to undergo backprop during training? Won't the external-model's embeddings just be noise since they don't share embedding space with the model that is being trained?

If the external-model also undergoes training along with the model then I think that might work.

I've pondered a bit more about it and I was the one who was mistaken. I think the author made great observations. It's just that I don't want to go back to non token thinking. I don't want there to be a 13x13xE final output from the CNN. I really want there to be a visual vocabulary from which tokens are chosen. And I want this visual vocabulary to be fixed/untrainable/deterministic. That'd be very cool.

But why only choose 13x13 + 1? :(

I'm willing to bet that the author's conclusion of embeddings coming from CNNs is wrong. However, I cannot get the 13x13 + 1 observation out my head though. He's definitely hit on something there. I'm with them that there is very likely a CNN involved. And I'm going to put my bet on the final filters and kernel are the visual vocabulary.

And how do you go from 50k convolutional kernels (think tokens) to always 170 chosen tokens for any image? I don't know...

I'm probably wrong but the author may have have misunderstood input embeddings. Input embeddings are just dictionary lookup tables. The tokenizer generates tokens and for each token you find its embedding from the lookup.

The author is speculating about an embedding model but in reality they're speculating about the image-tokenizer.

If I'm not wrong the text tokenizer Tiktoken has a dictionary size of 50k. The image tokenizer could have a very large dictionary size or a very small dictionary size. The 170 tokens this image tokenizer generates might actually have repeating tokens!

EDIT: PS. What I meant to say was that input embeddings do not come from another trained model. Tokens come from other trained models. The input embedding matrix undergoes back propagation (learning). This is very important. This allows the model to move the embeddings of the tokens together or apart as it sees fit. If you use embeddings from another model as input embeddings, you're basically adding noise.

I don't need to be convinced of their viewpoint, I live it. I live in another European country, in a city with over million people. I live midway between the middle of the city and the outskirts. Practially everything one needs is available within a 10 minute bike ride. The weather is atrocious sometimes but everyone deals with it by wearing a rain poncho. People who can't bike are on mobility scooters and e-wheelchairs (paid for by health insurance) who share the bike lane. Sometimes I'm too lazy and I take buses, trams, subways and trains which are ubiquitous. And I am not an avid cyclist, wish I was. My bike for me is a tool, a $150 second-hand tool.

Is that the same everywhere in the country or even the city? No. But it's common enough that you'll see parents carrying their kids around in cargo bikes practically anywhere in the country.

P.S. I was generalizing when I said everything's within a 10 minute bike ride to account for all people who live along the same radius as I do. For me personally, everything's available within a 5 minute *walk* - grocery stores, cafes, restaurants, bars, gym, park, playgrounds, kindergartens, schools, train station, doctor, dentist.

I don't refute that things can be poorly designed in the Netherlands or in Europe in general. It is not a utopia. I'll give it a 6/10, okay on-average, can do better.

The thing is that US cities and suburbs are much worse. Let me quote the title of article

> Walking the Netherlands: From Amsterdam to The Hague

Is walking something you can do in the US except in some rare exceptions? There are suburbs in the US i.e. residential ONLY places where there aren't any side walks. And the closest grocery store may be 10-15 minutes away by car. Doesn't that feel suffocating to have a home but not be allowed to walk out on your own two feet?

Nope, I disagree. It should be the first thing you watch before you move onto more technical videos. The videos are pleasant to watch and are made by a Canadian who moved to the Netherlands. They've got excellent A and B roll footage that truly shows you the contrast between American/Canadian and European cities.

And it should be emotional because a commute isn't just about efficiency. One spends a sizeable portion of their life commuting. The human aspect of it can't be ignored. The depressing feeling you get while standing in the middle of a Walmart parking lot is real. At least it was for me.

I think everyone should watch the Not Just Bikes YouTube channel especially their series on Strong Towns [1]. It shows just how bad parking mandates are and just how bad American car centric city planning is. As they article pointed out, you need such vanguard cities to pave the way for better cities, with less parking and fingers crossed, fewer cars.

[1] https://youtu.be/y_SXXTBypIg?si=N8PF71wMhVv6aH9G

> People with ADHD operate on an “interest-based nervous system” so they’re motivated by urgency

Was the decision to have a 3 day trial instead of something longer have anything do with creating urgency and capitalizing on the impulsivity of people with ADHD?

I imagine people with ADHD are more likely to be excited by tools that help and they can often mistake the initial excitement with the tool actually helping manage their executive dysfunction. It's hard for me to see how anyone with ADHD will be able to judge this tool as being truly good for them within 3 days.

I have a question for people who know how American housing works. Are property taxes based on current price? If so, who sets current price? Also does property tax fill the coffers of local municipalities or the state?

I get your point but in your analogy don't you think startups should be considered amateur chess players rather than experts? "Forced errors" are a thing in chess and tennis but I don't see that applying to startups. I can see Apple, Microsoft and other big companies playing chess through lobbyists and lawsuits but I think the article was about startups.

P.S. I've never founded a company or have any experience with it.

Not denying your point but in some applications (which aren't outliers) a good cache can increase number of requests served 100x. Achieving those with read-replicas is easy and possible but can be quite expensive.

A read-replica doesn't give you perfect transactional guarantees as well. Read operations to the replica after a write operation to the leader might still give you stale data (lagging by a few milliseconds).

The difference I see is that in this case, it's the same middleman that visits all the stores (and is welcome at the store) and it's the same middleman that is making the price recommendations to all the stores at near real-time.

The company that I work for uses a dynamic pricing SaaS. This SaaS basically scrapes the 10-15 competitors' ecommerce websites, including the website of the company that I work for and they provide an excel sheet with every product and the price at which every company sells it for. Additionally if you have a contract with this company, you can also provide it your product stock information on a daily basis. This tells the SaaS how quickly or slowly some products are being sold.

It provides two services that I could tell (I'm not a pricing expert)

1. It gives you a "dumb" daily recommendation per product to either lower or increase the price so that your price in on par with the competitor who is selling it for the highest price.

2. It also gives you a "smart" daily recommendation per product to either lower or increase the price so that your price in on par with the competitor who is making the most sales.

Now I see no reason why all 10-15 competitors can't be subscribed to this SaaS and therefore all competitors are pricing their products practically the same. It's hard for me to come to terms with it, because on one hand it is "pricing it right" and on another it is algorithmic collusion.

The only moral issue I see here is that some products are "essential" (think food, toiletries) and therefore the customer will buy these products at prices that they are not comfortable with.

My bad, not pop ups, instead there's an info symbol next to each option under "advanced details". I was being lazy with the terminology, I don't know what such info buttons are called.

That note did not show up for me when launching the EC2 instance. Probably because I was switching between ARM and AMD quite a bit which caused the info pop ups to stop refreshing and not display info for T4g and were stuck on T2 and T3.

Thanks for pointing that out, I never noticed that setting. I just tried to launch an EC2 instance looking for this setting and here's what I found.

> A credit specification is only available for T2, T3, and T3a instances. Selecting Unlimited for the credit specification allows applications to burst beyond the baseline for as long as needed at any time.

So that would mean Unlimited is not a setting available for T4g (ARM instance) and therefore may explain inconsistent behavior in the ARM instance.

Aren't 't' instances burst instances? They need to be under constant load for a long time before their burst credits for CPU, memory, network and EBS run out, after which they fall back on their baseline performance.

> It does appear that the Arm-based instances can’t consistently maintain the same performance at high request rates.

I'm unwilling to trust that statement at face value for now given it's been tested against a 't' instance.

EDIT: Removed note about network burst credits in compute and memory optimized instances. I'm not sure if these instances have that.

In my opinion that is not how you should be timestamps localized to a time zone in Postgres. You should instead use:

  SELECT now()::timestampz at time zone 'US/Pacific'

The output is a timestamp that has no time zone offsets and therefore is localized. However I've noticed most novice programmers don't fully understand the implications of 'at time zone ...' syntax.

It makes sense to me that the actual time zone is not stored within the timestamp itself since tomorrow a time zone can introduce DST and therefore some timestamps that were valid today will no longer be valid tomorrow.

I guess it's just me but that website is quite readable on desktop. I mean, all I'm seeing is perfectly legible prose and satire while still being a legitimate config file. I literally would share this website as proof of Yaml's readability.

As to your frustration with logic and code being stuffed into YAML, I totally get it. Everyone coming up with their DSL, I'm guilty of it too. I think YAML unfairly takes the blame for unnecessary abstractions built on top of it. YAMLs only fault being, that you actually can write DSLs with it easier than with anything else.

I've always wondered, what about tooth brushes? Don't they release microplastics too?

Woah, that's news to me. Is that true even if triggers are used to update a column?

  CREATE OR REPLACE FUNCTION
      update_updated_at_function()
  RETURNS TRIGGER AS $$
  BEGIN
      NEW.updated_at = now();
      RETURN NEW;
  END;
  $$ language 'plpgsql';


  CREATE TRIGGER
      update_updated_at_trigger
  BEFORE INSERT OR UPDATE ON
      "my_schema"."my_table"
  FOR EACH ROW EXECUTE PROCEDURE
      update_updated_at_function();
  END $$;

Is it possible for two rows to have `updated_at` timestamps that are different from the transaction commit order even if the above function and trigger are used? It's alright if `updated_at` and the commit timestamp are not the same, but the `updated_at` must represent commit order accurate to the millisecond/microsecond.

I hope this starts a trend of digitizing many aspects of the German bureaucracy, official or otherwise.

I've sitting on my hands for the last three months to fill out some paperwork. Additionally, I have to apply for this specific appointment via email a few months in advance and it takes weeks for them to reply and if there's any back and forth, that's a few more weeks.

I love Germany, love the bureaucracy too, I just hate the lack of digitization.

You could approach data quality testing as if you're testing another piece of software by writing tests. We use dbt and it makes writing tests against models (think tables in a db) very easy.

For example, if you have a regional_orders table. You write tests in SQL to test your assumptions about that data:

* I expect regional_orders table to contain no duplicates entries.

* I expect regional_orders to ship to only a specific region.

* So on...

This has worked fairly well so far for me. But are these kinds of tests sufficient? Am I missing something?

I think the Personal Liability insurance that almost everyone has in Germany is an overlooked societal lubricant. For €10 a month, you get €10 million in coverage for accidental damage.

* No incentive for revenge law suits.

* Damages are well regulated so you can't sue for €1 million just because someone's bicycle ran over your toe.

* Can't sue for exorbitant medical expenses because in most cases there are no personal medical expenses i.e. covered anyway by govt health insurance.

I'm unsure if personal liability insurance is popular elsewhere in Europe.

Oof that stings. I bought a Fairphone 4 couple of months earlier after much deliberation about whether to wait for Fairphone 5 or not.

I seemed to have missed out on:

* +22% increased CPU GHz

* ~40% increased performance on GPU benchmarks at same power consumption

* D̶u̶a̶l̶ ̶S̶i̶m̶

* +7% battery

* IP54 -> IP55 water resistance?

* And almost every other metric that I don't care much about e.g. Network, Display and Camera.

What's the F4 better at?

* Slightly smaller display 6.30 inch vs 6.46 inch

* ̶W̶e̶i̶g̶h̶t̶!̶ ̶I̶ ̶b̶e̶t̶ ̶t̶h̶e̶ ̶F̶5̶ ̶i̶s̶ ̶h̶e̶a̶v̶i̶e̶r̶ ̶t̶h̶a̶n̶ ̶F̶4̶ ̶(̶2̶2̶5̶ ̶g̶r̶a̶m̶s̶)̶ ̶b̶u̶t̶ ̶I̶ ̶s̶e̶e̶ ̶n̶o̶ ̶w̶e̶i̶g̶h̶t̶ ̶m̶e̶n̶t̶i̶o̶n̶e̶d̶ ̶f̶o̶r̶ ̶F̶5̶.̶

* No change in Audio jacks, both seem to still have a single USB C connector.

I guess I'll just suck it up and wait for the Fairphone 6 or potentially Fairphone 7 in 5 years.

EDIT: Two corrections. I wasn't aware F4 also has dual sim. Apparently F5 is lighter at 212 grams.

Here's something to add some fuel to your philosophical fire :)

https://en.m.wikipedia.org/wiki/China_brain

Any typical analytics use case, where a single dataset (think a single table) is less than 2TB and the total data size is less than 10-20 TB in the organization. If that's you, you very likely don't need Snowflake and definitely not Databricks. An RDBMS would suffice, I recommend Postgres.

These sizes are just my thumb rules. Your mileage may vary.

I have people in my network use Snowflake with a total data size of 50GB!!! It makes my blood boil.

If you've ever tried a pair of Gunnar ambers you'd know what real blue light filtering is. With them on, it's not a subtle effect. You see with an amber tint. The monitor just appears less noisy. I've worked 12 hours straight in them in the past after which I'm physically and mentally tired but my eyes aren't. Without them I can't work, which is exactly how you know they're working.

If you're a programmer just looking at text all day, they are worth every penny. If you're a designer and color perception is important to you, you're out of luck.

Not affiliated to Gunnar in any way. Just a happy customer of 6 years.

Does anyone know how Mountpoint compares to s3fs-fuse in terms of performance and features?

I'm a resident of Germany and I visited the US this year. I'm quite conscious of my 5 year old's sugar intake so I almost obsessively look at the labels looking for the percentage of sugar a product contains at grocery stores.

Reading the labels in the US was __extremely aggravating__ due to the serving size bullshit and it gave me a headache after calculating the percentage of sugar in the products after like the 10th product. Here in Germany, the label states the number of grams of sugar for every 100 grams of product, so you don't even have to calculate the percentage. Comparing products in the grocery store aisles is really easy in Germany.